run.py

import wandb
import d3rlpy
import argparse
import traceback

from d3rlpy.dataset import ReplayBuffer, InfiniteBuffer
from d3rlpy.preprocessing import StandardObservationScaler
from d3rlpy.logging import CombineAdapterFactory, FileAdapterFactory, TensorboardAdapterFactory
from fcev import FCEVEnv, load_drive_cycle
from d3rlpy.algos import (
    TD3PlusBCConfig, IQLConfig, CQLConfig, BCQConfig,
    CalQLConfig, AWACConfig, ReBRACConfig, TACRConfig,
    PLASConfig, PRDCConfig, BEARConfig
)

from typing import Any, Optional
from d3rlpy.logging import WanDBAdapter
from d3rlpy.logging.logger import (
    AlgProtocol,
    LoggerAdapter,
    LoggerAdapterFactory,
    SaveProtocol,
)

# ---------- WandB Logger Factory ----------
class GWanDBAdapterFactory(LoggerAdapterFactory):
    r"""WandB Logger Adapter Factory class.

    This class creates instances of the WandB Logger Adapter for experiment
    tracking.

    Args:
        project (Optional[str], optional): The name of the WandB project.
            Defaults to None.
    """

    _project: Optional[str]

    def __init__(self, project: Optional[str] = None, experiment_name: Optional[str] = None,) -> None:
        self._project = project

    def create(
        self, algo: AlgProtocol, experiment_name: str, n_steps_per_epoch: int
    ) -> LoggerAdapter:
        return WanDBAdapter(
            algo=algo,
            experiment_name=experiment_name,
            n_steps_per_epoch=n_steps_per_epoch,
            project=self._project,
        )



# ---------- Algorithm Config Dictionary ----------
def get_algo_configs():
    # Algorithm configurations with encoder and observation preprocessing settings
    algo_configs = {
        "TD3PlusBC": TD3PlusBCConfig(actor_encoder_factory=d3rlpy.models.DefaultEncoderFactory(use_batch_norm=True, dropout_rate=0.2), critic_encoder_factory=d3rlpy.models.DefaultEncoderFactory(use_batch_norm=True, dropout_rate=0.2), observation_scaler=StandardObservationScaler()),
        "IQL": IQLConfig(actor_encoder_factory=d3rlpy.models.DefaultEncoderFactory(use_batch_norm=True, dropout_rate=0.2), critic_encoder_factory=d3rlpy.models.DefaultEncoderFactory(use_batch_norm=True, dropout_rate=0.2), observation_scaler=StandardObservationScaler()),
        "CQL": CQLConfig(actor_encoder_factory=d3rlpy.models.DefaultEncoderFactory(use_batch_norm=True, dropout_rate=0.2), critic_encoder_factory=d3rlpy.models.DefaultEncoderFactory(use_batch_norm=True, dropout_rate=0.2), observation_scaler=StandardObservationScaler()),
        "BCQ": BCQConfig(actor_encoder_factory=d3rlpy.models.DefaultEncoderFactory(use_batch_norm=True, dropout_rate=0.2), critic_encoder_factory=d3rlpy.models.DefaultEncoderFactory(use_batch_norm=True, dropout_rate=0.2), observation_scaler=StandardObservationScaler()),
        "CalQL": CalQLConfig(actor_encoder_factory=d3rlpy.models.DefaultEncoderFactory(use_batch_norm=True, dropout_rate=0.2), critic_encoder_factory=d3rlpy.models.DefaultEncoderFactory(use_batch_norm=True, dropout_rate=0.2), observation_scaler=StandardObservationScaler()),
        "AWAC": AWACConfig(actor_encoder_factory=d3rlpy.models.DefaultEncoderFactory(use_batch_norm=True, dropout_rate=0.2), critic_encoder_factory=d3rlpy.models.DefaultEncoderFactory(use_batch_norm=True, dropout_rate=0.2), observation_scaler=StandardObservationScaler()),
        "ReBRAC": ReBRACConfig(actor_encoder_factory=d3rlpy.models.DefaultEncoderFactory(use_batch_norm=True, dropout_rate=0.2), critic_encoder_factory=d3rlpy.models.DefaultEncoderFactory(use_batch_norm=True, dropout_rate=0.2),  q_func_factory=d3rlpy.models.QRQFunctionFactory()),
        "TACR": TACRConfig(actor_encoder_factory=d3rlpy.models.DefaultEncoderFactory(use_batch_norm=True, dropout_rate=0.2), critic_encoder_factory=d3rlpy.models.DefaultEncoderFactory(use_batch_norm=True, dropout_rate=0.2), observation_scaler=StandardObservationScaler()),
        "PLAS": PLASConfig(actor_encoder_factory=d3rlpy.models.DefaultEncoderFactory(use_batch_norm=True, dropout_rate=0.2), critic_encoder_factory=d3rlpy.models.DefaultEncoderFactory(use_batch_norm=True, dropout_rate=0.2), observation_scaler=StandardObservationScaler()),
        "PRDC": PRDCConfig(actor_encoder_factory=d3rlpy.models.DefaultEncoderFactory(use_batch_norm=True, dropout_rate=0.2), critic_encoder_factory=d3rlpy.models.DefaultEncoderFactory(use_batch_norm=True, dropout_rate=0.2), observation_scaler=StandardObservationScaler()),
        "BEAR": BEARConfig(actor_encoder_factory=d3rlpy.models.DefaultEncoderFactory(use_batch_norm=True, dropout_rate=0.2), critic_encoder_factory=d3rlpy.models.DefaultEncoderFactory(use_batch_norm=True, dropout_rate=0.2), observation_scaler=StandardObservationScaler()),
    }
    return algo_configs

# ---------- Training Function ----------
def train(args):
    algo_configs = get_algo_configs()

    if args.algo not in algo_configs:
        raise ValueError(f"Unsupported algorithm: {args.algo}")

    # Load dataset
    with open(args.dataset_path, "rb") as f:
        dataset = ReplayBuffer.load(f, InfiniteBuffer())

    # Load environment for evaluation
    env = FCEVEnv(load_drive_cycle(args.drive_cycle))

    config = algo_configs[args.algo]
    algo = config.create(device=args.device)

    # Setup logger
    logger_adapters = [
        FileAdapterFactory(root_dir=f"d3rlpy_logs/{args.algo}"),
        TensorboardAdapterFactory(root_dir=f"tensorboard_logs/{args.algo}")
    ]
    if args.wandb:
        logger_adapters.append(GWanDBAdapterFactory(experiment_name=f"{args.algo}-run", project=args.wandb_project))

    logger_adapter = CombineAdapterFactory(logger_adapters)

    try:
        print(f"\n🚀 Starting training: {args.algo}")

        algo.fit(
            dataset,
            n_steps=args.n_steps,
            n_steps_per_epoch=args.n_steps_per_epoch,
            logger_adapter=logger_adapter,
            evaluators={
                'init_value': d3rlpy.metrics.InitialStateValueEstimationEvaluator(),
                'soft_opc': d3rlpy.metrics.SoftOPCEvaluator(return_threshold=100),
                'action': d3rlpy.metrics.ContinuousActionDiffEvaluator(),
                'environment': d3rlpy.metrics.EnvironmentEvaluator(env),
                'Advantage': d3rlpy.metrics.DiscountedSumOfAdvantageEvaluator()
            },
        )

        print(f"\n✅ Training finished for: {args.algo}")

    except Exception as e:
        print(f"\n❌ Training failed: {args.algo}")
        print(traceback.format_exc())
        wandb.finish()


# ---------- Main CLI ----------
if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="Offline RL training for FCEV")

    parser.add_argument("--algo", type=str, default="AWAC",
                        choices=list(get_algo_configs().keys()),
                        help="Name of the offline RL algorithm")
    parser.add_argument("--dataset-path", type=str, default="datasets/fcev-mpc-v1.h5",
                        help="Path to the .h5 dataset file")
    parser.add_argument("--drive-cycle", type=str, default="CLTC-P-PartI.csv",
                        help="Path to the drive cycle CSV file")
    parser.add_argument("--n-steps", type=int, default=10000,
                        help="Total number of training steps")
    parser.add_argument("--n-steps-per-epoch", type=int, default=100,
                        help="Steps per epoch")
    parser.add_argument("--device", type=str, default="cuda:0",
                        help="Training device (e.g., 'cpu', 'cuda:0')")
    parser.add_argument("--wandb", action="store_true",
                        help="Enable WandB logging")
    parser.add_argument("--wandb-project", type=str, default="fcev-offline-benchmark",
                        help="WandB project name (used only if --wandb is enabled)")

    args = parser.parse_args()
    train(args)